Meta released an early technical overview of Scam Alert, an optional on-device tool that flags likely scam messages on WhatsApp, on August 12. All detection stays on the user’s phone, and no message content reaches Meta or WhatsApp. The launch comes as India’s telecom regulator weighs bringing over-the-top (OTT) platforms under network-level spam rules, while the government pushes WhatsApp to defend other features on fraud-related grounds.

Why it matters: The Telecom Regulatory Authority of India (TRAI) is finalising the Telecom Commercial Communications Customer Preference Regulations (TCCCPR) Third Amendment. Telecom operators want the rules extended to WhatsApp, Signal and Telegram. The draft would require call and messaging apps to share their spam reports with telecom operators’ Do Not Disturb registry. It would also allow TRAI to invoke the Information Technology (IT) Act against non-compliant apps. Repeated non-compliance could put their Section 79 safe harbour at risk.

The Internet and Mobile Association of India (IAMAI) has opposed the draft as regulatory overreach. It argued that requiring OTT platforms to share data with access providers amounts to “unconstitutional expropriation of valuable proprietary data.”

Scam Alert cuts against that data-sharing model. Telecom spam control depends on network-level visibility into who is messaging whom. Scam Alert keeps that visibility off the network:

  • The classification model runs on the device, not on Meta’s servers.
  • No message content leaves the phone for classification.
  • Nothing reaches Meta’s servers unless the user reports a message.

The tool shows that a platform can fight scams while preserving end-to-end encryption, without the spam-report sharing that the draft would require.

How Scam Alert works: Once a user turns it on, WhatsApp downloads a machine learning model to the device. The model classifies whether incoming messages from non-contacts match known scam patterns. It draws on scam conversations that users have voluntarily reported. Meta says the model is small enough to run on mobile hardware and simple enough to be published for independent review. This allows the system to work without server-side classification.

  • A flagged message triggers a warning in the chat that is visible only to the user.
  • The user can block, report or continue the conversation.
  • Marking a chat as trusted removes the warning and stops future flags on it.
  • A user who trusts a chat can opt to share the last five messages with WhatsApp to improve accuracy.

What Meta does and does not see: Meta receives two categories of anonymous, aggregate data, and only after differential privacy noise is added. Warning counts show how often the model flagged a scam. User action counts show whether users trusted or reported flagged chats. Meta says these metrics measure whether the feature works without revealing an individual’s messages, contacts or conversations.

The telemetry passes through a confidential computing pipeline built on Trusted Execution Environments (TEEs). Meta says the TEE blocks shell access at runtime from its own staff and networked systems. This means no engineer can read unaggregated data.

The verifiability claims: Meta built three transparency guarantees into the system, each of which it says security researchers can independently verify:

  • No targeted model delivery: Every model version, including experimental ones, is published on a public append-only ledger before deployment.
  • Verifiable model behaviour: Meta publishes model weights so researchers can confirm that the model targets scams only.
  • User visibility: Users can review which messages were scanned through Account, Request Info and Scam Alert Activity.

Several mechanisms support the no-targeting claim. Model download requests route through an Oblivious HTTP (OHTTP) relay, which strips the user’s IP address. They also use anonymous credentials that prove the request comes from a real WhatsApp client without identifying which one. Experiment group assignment happens on the device using locally generated randomness. This means Meta’s servers cannot steer a user to a specific model variant.

Before loading a model, the client verifies a Cloudflare signature on the model manifest. It also cross-checks the manifest hash against the ledger. The client refuses to load the model if any step fails. Cloudflare, not Meta, holds the signing key. Meta is expanding its Bug Bounty programme to cover the pipeline and the model.

The encryption backdrop: The design comes amid a series of clashes between India and Meta over WhatsApp and fraud. In July, the Ministry of Electronics and Information Technology (MeitY) forced WhatsApp to halt its username feature in its largest market. The government argued that allowing users to message without revealing their phone numbers would increase phishing, impersonation and “digital arrest” scams.

Scam Alert takes a different approach to the same fraud concerns. It uses on-device detection instead of identity disclosure or network-level access sought by regulators. Meta published the overview alongside a limited beta, not a finished system. The company said it will continue stress-testing the architecture through its bug bounty community before a production release.

Also read: